1. The Quiet Risk
Most worry about AI concerns what it might do wrong in a single answer. A slower-moving risk gets less attention. The system you rely on may change underneath you, and you may have no say in when or how.
A hosted model is operated by a vendor. The vendor may update it, adjust its behaviour, change its limits, alter its price or retire it. Each of those changes is reasonable from the vendor's side. From a client's side, each can break something that worked.
2. What a Subscription Grants
A subscription grants access. It does not grant ownership of the weights, the logs, or the terms on which both may change.
- The weights are the model itself. The client uses them but does not hold them.
- The logs record what was asked and answered. The vendor decides what is kept, for how long and for what use.
- The terms govern both, and the vendor can revise them.
Access lasts as long as the vendor chooses to provide it on terms the client accepts. Everything built on that access carries the same condition.
A subscription answers the question "can I use it today?" It does not answer "will it behave the same way next year, and will it exist?"
3. How Drift Shows Up
For a family office or executive team, drift is rarely dramatic. It appears as small changes that accumulate.
- Style and tone shift. A drafting assistant tuned to a principal's voice starts to write differently.
- Behaviour on edge cases changes. A question that was answered is now declined, or the reverse.
- A workflow quietly degrades. A chain of steps calibrated against one model's habits gets worse with a new one.
- A model disappears. The version a team relied on is retired, and the replacement is not identical.
In each case, someone has to notice, diagnose and rework. The cost falls on the client, and the timing was set by the vendor.
4. What Holding the Model Changes
When the model runs on hardware the client owns, change happens when the client decides. The weights on the appliance do not alter overnight because a vendor updated something remotely. A client who has refined their workflows around a model can keep using it.
That is a statement about control, not about quality. A locally held model can also be replaced, upgraded or supplemented, and the client is involved in those decisions. SARION Appliances offer the standard sovereign model line-up, and at the higher tiers frontier-scale models, with parameters counted in the trillions, are available as options. The choice of what runs, and when it changes, remains the client's.
5. An Asset, Not an Access Right
Ownership changes the nature of what the client holds. The system becomes an asset on the client's own premises, acquired once, rather than a right of access that someone else grants and can revise. Using it more, relying on it more, building more of the office's work around it: none of that changes the client's relationship with anyone else.
Ownership also brings the obligations of any serious asset: a proper place to house it, power, service and, in time, renewal. A family that already maintains estates, vessels and collections will recognise the pattern. The asset is theirs, and so is its care.
6. Questions for Any AI Vendor
- Can the vendor change or withdraw the model I rely on, and with how much notice?
- Who holds the logs of my usage, for how long, and for what purposes?
- Can the terms change, and what are my options when they do?
- What happens to my workflows if the product is discontinued or the vendor is acquired?
- Can I keep using the system at all if I disagree with a change?
7. What Stability Does Not Mean
A held model is not a frozen one. Models improve, and a client will want to benefit. The difference is that the client chooses when to move, and can test a new model against their own work before relying on it.
Stability also does not mean the model is correct. Verification of outputs remains necessary under any model.
8. A Short Test
Imagine your vendor announces tomorrow that a model you rely on will be retired in 90 days. Ask how much of your work depends on its specific behaviour, who would rework it, and who would pay. For a client with the model on their own hardware, the announcement is somebody else's news.
To discuss model choices and continuity for your own deployment, contact us at volta.quest.